A method and system for dissertation review

Abstract

The present application relates to the technical field of text data processing, and more particularly to a dissertation review method and system, comprising obtaining a dissertation document to be reviewed, parsing the document into structured data objects, and distributing data blocks in the structured data objects to multiple preset review domains according to content semantics; for at least one core review domain, extracting technical entities in the data blocks, dynamically dividing the technical entities using a community discovery algorithm, forming multiple technical communities, instantiating at least one review intelligent agent for each technical community, and constructing a parallel review computation graph; effectively solving the distortion of long text processing and the bottleneck problem of manual review, and introducing a two-stage adversarial game architecture, through a double-checking mechanism of local fact verification and global logic alignment, significantly suppressing the "fact illusion" of large models, and eliminating the logical conflicts between multi-agent reviews, greatly improving the accuracy and logical consistency of the automated review results.

Description

Technical Field

[0001] This invention relates to the technical field of text data processing, and in particular to a method and system for reviewing dissertations. Background Technology

[0002] Dissertations are the ultimate measure of a graduate student's academic level, innovation ability, and training quality, and a core component of the higher education quality assurance system. With the popularization of higher education and the continuous expansion of graduate student enrollment, the number of master's and doctoral dissertations awaiting review is growing exponentially. Statistics show that the number of dissertations requiring review in China each year has reached the millions. Faced with such a massive amount of data, the traditional manual review model is facing severe challenges, while existing automated review technologies have significant technical bottlenecks.

[0003] First, traditional manual review methods struggle to balance efficiency, quality, and fairness. Qualified domain experts are a scarce resource. During peak graduation season, a single supervisor often has to review dozens of papers in a very short time, easily leading to "review fatigue" and overlooking details such as data fabrication and logical flaws. Second, different experts have vastly different academic tastes and standards of leniency, making subjective bias difficult to avoid. Finally, modern dissertations (especially in STEM fields) are often highly interdisciplinary, containing multiple complex technology stacks (for example, a single paper may involve deep learning algorithms, blockchain architecture, and control theory simultaneously). It is difficult for a single expert to master every innovative point in a paper, easily leading to perfunctory reviews or misjudgments.

[0004] Secondly, while existing automated review technologies have undergone several generations of development, they have not yet completely solved the core pain points. First-generation technologies primarily rely on rule-based, shallow checks, using regular expressions and statistical models to check for typos, formatting errors, and proper citation of references. Their limitation lies in their inability to understand the deeper semantics of the text, rendering them ineffective in addressing core academic questions such as "whether the experimental method design is reasonable" and "whether the conclusions effectively support the hypothesis."

[0005] The second-generation technology is an end-to-end generation technique based on monolithic large models. With the emergence of large models such as ChatGPT and Claude, papers can be directly input into these models to generate reviews. However, this technology faces three major challenges: First, the "Lost-in-the-Middle" phenomenon. Dissertations are extremely long (typically 30,000 to 100,000 words), and monolithic large models, when dealing with such lengthy texts, often focus their attention mechanisms on the beginning and end, neglecting crucial argumentation processes and experimental details in the middle. Second, the static attention allocation problem. Monolithic large models use a uniform attention mechanism across the entire text, making it difficult to distinguish between "core innovations" and "background information." For complex papers containing multiple independent experiments or methods, large models often cannot perform fine-grained differentiation in their reviews, resulting in vague and general evaluations. Third, the "Hallucination" problem. When lacking external knowledge support, large models are prone to fabricating non-existent experimental data, fabricating references, or even incorrectly pointing out flaws not present in the text. Such "serious nonsense" is unacceptable in serious academic reviews.

[0006] Third-generation technologies attempt to use early multi-agent slice review, employing a "divide and conquer" strategy to distribute paper slices to different agents. Their core flaw lies in "logical inconsistency." Because each agent works independently, there is a lack of a global logical alignment mechanism, often resulting in contradictory evaluations: "the method review agent considers the algorithm design ingenious, while the experiment review agent considers the results unreliable." The system lacks a mechanism to determine whether the problem stems from the algorithm itself, poor experimentation, or a misinterpretation of the algorithm by the experimental agent. This logical mutual exclusion cannot be automatically resolved in existing systems.

[0007] In summary, existing technologies generally suffer from shortcomings when processing extremely long and interdisciplinary dissertations, such as incomplete review coverage, insufficient analytical depth, difficulty in ensuring factual accuracy, and logical conflicts in multi-source review opinions. Summary of the Invention

[0008] The technical problem solved by this invention is that existing technologies generally suffer from defects such as incomplete review coverage, insufficient analytical depth, difficulty in ensuring factual accuracy, and logical conflicts in multi-source review opinions when dealing with extremely long and interdisciplinary dissertations.

[0009] To solve the above-mentioned technical problems, the present invention provides the following technical solution: First aspect, a method for reviewing dissertations, comprising:

[0010] Step S1: Obtain the thesis document to be reviewed, parse the document into a structured data object, and distribute the data blocks in the structured data object to multiple preset review domains according to the content semantics.

[0011] Step S2: For at least one core review domain, extract the technical entities from the data block, use the community discovery algorithm to dynamically divide the technical entities and form multiple technical communities, and instantiate at least one review agent for each technical community to construct a parallel review computation graph.

[0012] Step S3: Perform first-order local adversarial critique verification on the review opinions generated by each review agent. The first-order local adversarial critique verification includes fact checking and triggering a self-correction loop based on the confidence level of the fact checking.

[0013] Step S4: Perform a second-order global logical consistency game on all review opinions after the first-order local adversarial critique verification. The second-order global logical consistency game includes determining the logical relationship between each review opinion and triggering multiple rounds of adversarial negotiation until logical convergence when there is a logical conflict.

[0014] Step S5: Weighted fusion is performed based on the confidence level of each review opinion to generate the final review result.

[0015] As a preferred embodiment of the thesis review method described in this invention, the plurality of preset review domains include a reference domain. Abstract field Method domain Other integrated domains .

[0016] As a preferred embodiment of the thesis review method described in this invention, the community detection algorithm is the Louvain algorithm, and the instantiation is containerized instantiation.

[0017] As a preferred embodiment of the thesis review method described in this invention, the fact-checking calculates the confidence level using a fact consistency discriminant function, and triggers the self-correction loop when the confidence level is lower than a preset threshold.

[0018] As a preferred embodiment of the thesis review method described in this invention, the determination of the logical relationship between the review opinions is achieved by constructing a logical mutual implication matrix using a Natural Language Inference (NLI) model.

[0019] As a preferred embodiment of the thesis review method described in this invention, the weighted fusion adopts an exponential decay algorithm based on self-correcting rounds;

[0020] The weight of the review comments is negatively correlated with the number of times they are self-corrected.

[0021] Secondly, a thesis review system includes a document reconstruction and routing module, a dynamic graph construction module, a local verification module, a global alignment module, and a result fusion module.

[0022] The document reconstruction and routing module is used to obtain the thesis documents to be reviewed, parse the documents into structured data objects, and distribute the data blocks in the structured data objects to multiple preset review domains according to the content semantics.

[0023] The dynamic graph construction module is used to extract technical entities from the data block for at least one core review domain, dynamically divide the technical entities based on the community discovery algorithm to form multiple technical communities, and instantiate at least one review agent for each technical community to construct a parallel review computation graph.

[0024] The local verification module is used to perform first-order local adversarial critique verification on the initial review opinions generated by each review agent, including fact-checking and triggering a self-correction loop based on the confidence level of the fact-checking.

[0025] The global alignment module is used to perform a second-order global logical consistency game on all review opinions, including determining the logical relationship between each review opinion and triggering multiple rounds of adversarial negotiation until logical convergence when there is a logical conflict.

[0026] The result fusion module is used to perform weighted fusion based on the confidence level of each review opinion to generate the final review result.

[0027] As a preferred embodiment of the thesis review system described in this invention, the document reconstruction and routing module is specifically used to distribute the data block to the reference domain. Abstract field Method domain Other integrated domains .

[0028] The beneficial effects of this invention are as follows: By reconstructing documents and semantic routing, ultra-long text papers are transformed into structured data and distributed to orthogonal review domains. Combined with dynamic graph arrangement technology, comprehensive coverage and fine-grained analysis of the paper content are achieved, effectively solving the problems of distortion in long text processing and bottlenecks in manual review. Furthermore, a two-order adversarial game architecture is introduced, which significantly suppresses the "fact illusion" of large models and eliminates logical conflicts between multi-agent reviewers through a dual verification mechanism of local fact-checking and global logical alignment, greatly improving the accuracy and logical consistency of automated review results. Moreover, the use of a confidence decay algorithm based on self-correcting rounds for score fusion makes the review results more objective and robust, providing reliable technical support for the scientific evaluation of dissertations. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the basic process of a thesis review method provided in one embodiment of the present invention. Detailed Implementation

[0030] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0031] Example 1, referring to Figure 1 As an embodiment of the present invention, a thesis review method is provided, comprising:

[0032] Step S1: Obtain the thesis document to be reviewed, parse the document into a structured data object, and distribute the data blocks in the structured data object to multiple preset review domains according to the semantics of the content.

[0033] Specifically, in this embodiment, the thesis document to be reviewed is first received. This document includes text data in PDF or Word format. Then, an existing OCR (Optical Character Recognition) engine (such as PaddleOCR or Tesseract) is used to recognize the text content in the PDF. Combined with an existing layout analysis model (such as LayoutXLM or LayoutLM series models), the document's layout information is parsed. This layout information includes headings, paragraphs, figures, and formulas. The entire document is then reconstructed into a structured JSON object with a hierarchical index. In specific operations, [the following steps are used]. It represents a structured JSON object, and the JSON object retains the logical structure of the original document, which facilitates subsequent content location and semantic analysis.

[0034] After document reconstruction is complete, a semantic routing agent is invoked to map data blocks (such as chapters and paragraphs) in the JSON object into high-dimensional semantic vectors. (This routing agent integrates a pre-trained semantic feature extraction model (such as BERT-Large or RoBERTa)). Mapped to four orthogonal semantic review domains, including the reference domain. Abstract field Method domain Other integrated domains In one embodiment, by calculating the similarity between the data block and the semantic centers of each review domain, each data block is automatically distributed to the semantic review domain with the highest similarity, i.e., the most matching semantic review domain. Orthogonal partitioning ensures comprehensive review, systematically evaluating the paper from different dimensions. In one embodiment, the reference domain... Responsible for verifying the accuracy and relevance of citations, abstract field Used to assess the conciseness of content, method domain Used for in-depth analysis of core innovation points, and other comprehensive domains. This step is used to evaluate the overall contribution and writing quality of a paper. It transforms unstructured raw documents into machine-readable and analyzable structured data.

[0035] Step S2: For at least one core review domain, extract the technical entities from the data block, use the community discovery algorithm to dynamically divide the technical entities and form multiple technical communities, and instantiate at least one review agent for each technical community to construct a parallel review computation graph.

[0036] In this embodiment, specifically for the reference field... Abstract field Method domain Other integrated domains Instantiate the corresponding review agents for each method domain. The meta-analysis agent (Meta-Analyzer) is activated, and using Named Entity Recognition (NER) and clustering strategies, it parses out the contents of the paper. Each independent technological innovation point is used to dynamically construct a parallel review computation graph, which forms a dynamic graph. Its mathematical expression is:

[0037]

[0038] Among them, the node set Corresponding to the innovation points generated One method to evaluate the agent, edge set Semantic Dependency Connections between agents are used to define the flow path and collaborative topology of contextual information during the review process, ensuring that logically related technical points can share intermediate inference states during the review.

[0039] In one embodiment, existing Named Entity Recognition (NER) technology is used to identify entities from the method domain. Core technical entities, such as algorithm names, model architectures, and key technical indicators, are extracted from the data blocks. These technical entities constitute the technical graph of the paper, and community detection algorithms (such as the existing Louvain algorithm) are used to dynamically partition these technical entities, dividing the technical entity network into several technical communities with tight internal connections and sparse external connections. Each technical community typically corresponds to an independent innovation point or a closely related technical cluster in the paper. A containerized review agent is instantiated for each dynamically identified technical community, and all review agents together constitute a parallel review computation graph. The nodes in the review computation graph represent review agents, and the edges represent possible logical connections or data dependencies between agents, ensuring parallel and in-depth mining of multiple innovation points in the paper and effectively simulating the process of collaborative review by experts from multiple fields.

[0040] Step S3: Perform first-order local adversarial critique verification on the review opinions generated by each review agent. The first-order local adversarial critique verification includes fact checking and triggering a self-correction loop based on the confidence level of the fact checking.

[0041] In this specific embodiment, each review agent in the system preferably uses real-time retrieval via Google Scholar and Arxiv, which involves calling these two web retrieval interfaces. Each agent performs multi-hop inference based on the Chain of Thought (CoT) mechanism, combining the full-text evidence of the paper retrieved by RAG with the external state-of-the-art benchmarks obtained from web search to generate an initial review vector. In the first stage: after each review agent node in the fixed review agent and the dynamic method review agent, a dedicated local critique agent is connected to construct an "execution-critique" dual structure; the local critique agent, based on the fact consistency judgment model, evaluates the initial review vector generated by the preceding review agent. Performing atomic-level fact-checking, and triggering a self-correction loop if the check fails, helps improve the factual accuracy of each reviewer's comments and suppresses the "illusion" of the large model;

[0042] In one embodiment, a dedicated local critique agent is linked to each review agent. The core task of this critique agent is to perform atomic-level fact-checking, comparing key factual statements (such as cited literature, data, formulas, etc.) in the initial review opinion with a knowledge base or the original document. The fact-checking quantifies confidence through a fact consistency discriminant function, preferably an NLI-based binary classifier, which outputs a probability score of "consistent" or "inconsistent". When the confidence score is lower than a preset threshold (e.g., 0.85), a self-correction loop is triggered, feeding back the verification result to the original review agent, requiring it to regenerate or revise the review opinion. This process iterates step S3 multiple times until the confidence score meets the requirement or the preset maximum number of iterations is reached.

[0043] Step S4: Perform a second-order global logical consistency game on all review opinions after the first-order local adversarial critique verification. The second-order global logical consistency game includes determining the logical relationship between each review opinion and triggering multiple rounds of adversarial negotiation until logical convergence when there is a logical conflict.

[0044] In this embodiment, specifically regarding the method domain... All local components are introduced into a global summary critical agent through nodes to construct a logically mutually implied matrix. It detects whether there are semantic conflicts or logical paradoxes among the review opinions of each sub-item, thereby eliminating the problem of viewpoint fragmentation and logical conflict that may occur in the parallel review of multiple agents, and ensuring the logical self-consistency of the final review system.

[0045] In one embodiment, after all review agents have completed local fact-checking and output stable review opinions, all review opinions are collected, and a logical mutual implication matrix is ​​constructed. In some preferred embodiments, this matrix is ​​constructed using a Natural Language Inference (NLI) model. For any two review opinions A and B, the NLI model determines whether their logical relationship is "implication," "neutral," or "contradictory," and fills the result into the matrix, calculating the global conflict loss in the matrix. For example, the number of "contradictory" relationships in the matrix can be used as the loss value; if the loss value is greater than 0, it indicates that there is a logical conflict; at this time, the system triggers multiple rounds of adversarial negotiation; in each round of negotiation, review agents holding conflicting viewpoints are organized to engage in dialogue, stating their respective evidence and reasoning processes, and attempting to reach a consensus. This process continues until the global conflict loss is reduced to 0, i.e., logical convergence.

[0046] Step S5: Weighted fusion is performed based on the confidence level of each review opinion to generate the final review result.

[0047] In this specific embodiment, after the global logic detection passes, the formatted agent is activated, the review vectors of all semantic review domains are aggregated, the final score is calculated based on the dynamic confidence weighting algorithm, and a structured review report is generated.

[0048] Each review opinion is accompanied by a confidence level, which combines its performance in the first-order verification (such as the factual consistency score) and its stability in the second-order game (such as the number of times it is challenged). To quantify the contribution of opinions with different confidence levels, a weighted fusion algorithm is preferred.

[0049] In one embodiment, an exponential decay algorithm based on self-correcting rounds is employed, the mathematical expression of which is:

[0050] Score_final=Σ(Score_i*e^(-λ*n_i));

[0051] Wherein, Score_final represents the final score, Score_i represents the initial score of the i-th reviewer's comment, n_i represents the number of self-corrections triggered by the reviewer's comment in the first-order validation, and λ represents the preset decay coefficient (e.g., λ=0.5). The more self-corrections a reviewer's comment has, the lower its confidence level, and its weight w_i=e^(-λ*n_i) also decreases accordingly. The confidence suppression mechanism makes the review results more robust and scientific, and reduces the interference of low-quality comments on the final conclusion.

[0052] By reconstructing documents and semantic routing, ultra-long text papers are transformed into structured data and distributed to orthogonal review domains. Combined with dynamic graph arrangement technology, comprehensive coverage and fine-grained analysis of the paper content are achieved, effectively solving the problems of distortion in long text processing and bottlenecks in manual review. Furthermore, a two-order adversarial game architecture is introduced, which significantly suppresses the "fact illusion" of large models and eliminates logical conflicts between multi-agent reviewers through a dual verification mechanism of local fact-checking and global logical alignment, greatly improving the accuracy and logical consistency of automated review results. Moreover, a confidence decay algorithm based on self-correcting rounds is used for score fusion, making the review results more objective and robust, providing reliable technical support for the scientific evaluation of dissertations.

[0053] Example 2, which is based on the previous example, differs from the previous example in that:

[0054] First, the system addresses the information distortion issue caused by the unstructured nature of PDF documents. It receives PDF data streams of dissertations and employs a high-precision OCR engine (such as PaddleOCR or Tesseract) combined with a deep learning layout analysis model (such as LayoutXLM) to identify text paragraphs, formulas (LaTeX format conversion), charts, and their geometric coordinates (bounding boxes) within the document.

[0055] Secondly, the identified content is reconstructed into a structured JSON object containing a hierarchical index. The object is stored in a tree structure, with nodes containing chapter titles, paragraph content, reference relationships, and page number metadata to ensure information integrity.

[0056] Then, the semantic routing agent is invoked, and based on the attention distribution of a pre-trained language model (such as BERT-Large), it... Each text block in the document is mapped to four orthogonal semantic review domains: the reference domain. Abstract field Method domain Other integrated domains .

[0057] The semantic routing agent maps data blocks (such as chapters and paragraphs) in a JSON object to high-dimensional semantic vectors, the mathematical expression of which is:

[0058] For a sequence of arbitrary data blocks extracted from a structured JSON tree Calculate the conditional probability distribution of its belonging to each predefined semantic review domain, using the following formula:

[0059]

[0060] in, Represents any element in the semantic domain tag set, with a value range of 1. , This refers to a pre-trained deep learning semantic feature extractor (such as the encoding layer of a BERT or RoBERTa model) used to map discrete text symbols into high-dimensional, continuous, semantically dense vectors. ,in Indicates the embedding dimension. The trainable projective weight matrix of the classifier has dimensions of . ,in Represents the total number of semantic domain categories. The bias vector of the classifier has dimensions of . Used to adjust the intercept of the classification plane. This represents the normalized exponential function, used to map the linearly transformed logits to a probability distribution that sums to 1.

[0061] Based on the maximum a posteriori probability criterion Sequence of data blocks Automatically distribute the data block to the corresponding domain processing task queue, and synchronously retain the context reference index (including page number index, paragraph ID, and geometric coordinate border) of the data block in the original PDF to ensure the traceability of subsequent processing.

[0062] Step S2 targets the most complex and innovative method domain. This embodiment abandons the traditional fixed review mode and adopts a density-based clustering strategy for dynamic map construction.

[0063] Specifically, the Meta-Analyzer is activated to scan the method domain text using Named Entity Recognition (NER) technology and extract the set of core technical entities. (For example, "Transformer architecture", "Dice loss function", "adaptive optimizer", etc.).

[0064] Next, construct the entity co-occurrence matrix. Calculate the co-occurrence frequency of entities within the same context window. Utilize the Louvain community detection algorithm to group the sets... Divided into A highly modular community For example, the algorithm automatically clusters entities related to the model structure into communities. Cluster entities related to the experimental loss function into communities. .

[0065] Finally, for each community Dynamically instantiate a containerized intelligent agent (e.g., launched via Docker containers). Its system prompt is dynamically generated by the community's central entity (e.g., "You are a review expert defined by Transformer..."), thus constructing a parallel review computation graph. Among them, the node set Corresponding generation One method to evaluate the agent, edge set Semantic dependency connections between corresponding agents are used to define the flow path of contextual information during the review process.

[0066] Dynamic map construction based on density-based clustering strategies includes:

[0067] Step S21, Extract The core technology entity set in Construct an entity co-occurrence matrix ;

[0068] Step S22, use the Louvain community detection algorithm to... Divided into A highly modular community ;

[0069] Step S23, for each community Instantiate a containerized intelligent agent Its system prompt is dynamically generated by the community's central entity.

[0070] To break the knowledge closure of large models, this embodiment configures a retrieval augmentation generation (RAG) interface and a real-time Internet search interface for each review agent in the system (including agents in fixed domains and dynamic method domains).

[0071] The agent performs multi-hop reasoning based on the Chain of Thought (CoT) mechanism. When it is necessary to verify arguments, the agent automatically generates query vectors. And it searches in local vector databases and on the Internet.

[0072] The retrieval augmentation generation (RAG) algorithm in step S3 specifically includes:

[0073] A hybrid strategy combining semantically dense retrieval and keyword-sparse retrieval is used to calculate the relevance score. The specific mathematical expression is as follows:

[0074]

[0075] in, Indicates the current query With candidate document fragments The total score of the mixed correlation between them This represents the mixed weight hyperparameter, with a value range of [value missing]. This is used to adjust the contribution ratio of dense and sparse searches to the total score. This represents the cosine similarity function, used to measure the directional similarity between two vectors in the semantic space. Indicates query The high-dimensional semantically dense vector obtained after mapping by the pre-trained embedding model. Represents candidate data blocks High-dimensional semantically dense vectors pre-computed and stored in a vector database. This represents the sparse retrieval score calculated based on the Okapi BM25 algorithm, used to capture precise literal matching features and term frequency weights between query terms and document terms. This represents the preset correlation cutoff threshold.

[0076] Calculate the scores for all candidate segments. The agent only takes in data that meets the criteria. The entire text fragments are used to construct a chain of contextual evidence to support the generation of review opinions, thereby balancing the depth of semantic understanding with the accuracy of keyword matching.

[0077] The local adversarial critique verification in step S4 specifically includes:

[0078] Step S41, Construction of the Fact Consistency Judgment Model:

[0079] To eliminate the "illusion" of the monolithic model, a local critical agent is chained after each review node in the computation graph to construct an "execution-critique" dual structure.

[0080] Local critical agents construct fact consistency discriminant functions based on conditional probability generative models. Used to calculate review opinion vector Relative to the source document context and the search for the chain of evidence Confidence score of authenticity The calculation formula is:

[0081]

[0082] in, The range of values (Logarithmic probability space) The total length of the tokens representing the sequence of review comments. The parameter is The probability distribution of the discriminator neural network, This indicates that the target category label is "faithful to the original text". The review comments are in the [number]th [section]. The hidden state vector at each time step.

[0083] In one embodiment, a fact-checking threshold is set. (For example, -0.5). When When the system determines that the check has failed, it immediately activates a self-correcting loop. The local critical agent generates specific error correction instructions. (For example: "The accuracy data cited in sentence 3 does not match Table 2 in the original text"), driving the review agent to perform an update:

[0084] Step S42, Self-correction triggering and optimization mechanism:

[0085] Set fact-checking thresholds .when When the system determines that the check has failed, it immediately activates the Self-Correction Loop. At this point, the local critical agent generates corrective gradient feedback. Or natural language correction instructions The review agent will perform the following updates:

[0086]

[0087] in, This represents the revised review vector. This indicates specific error correction instructions generated by the critical agent (e.g., "The accuracy data cited should be 98.5% instead of 99%)". This represents the KL divergence penalty coefficient, used to prevent the corrected text from deviating too far from the original semantic style. The Kullback-Leibler divergence is used to measure the generated distribution. With prior distribution The differences between them are addressed by introducing a KL divergence penalty term to prevent the corrected text from deviating too far from the original semantic style.

[0088] The global logical consistency game in step S4 specifically includes the following sub-steps:

[0089] Step S543, construct the logical mutual implication matrix:

[0090] For method domain of Each intelligent agent is constructed with the following dimensions. Logical mutual implication matrix , where any element in the matrix Quantitative characterization of the first Review comments from individual agents With the Review comments from individual agents The logical reasoning relationship between them is calculated using the following model:

[0091]

[0092] in, Representation matrix The Middle Line number The elements of the column belong to a discrete label set. , This represents a deep neural network model for natural language inference pre-trained based on the Transformer architecture, used to determine the logical inference relationship between two text segments. This represents the precondition input for the inference task, mapped here to the source node agent. Generated review text , The input of the assumptions for the reasoning task is represented here and mapped to the target node agent. Generated review text .

[0093] Step S44, global conflict loss calculation:

[0094] Based on matrix Define the global conflict loss function of the system. This is used to quantify the degree of logical inconsistency in current multi-agent systems.

[0095]

[0096] in, This represents a scalar value indicating the global conflict loss. Its value is always greater than or equal to 0. The larger the value, the more severe the internal logical conflicts within the system. This represents a double traversal and summation over all possible combinations of agent pairs in the matrix. This represents a logical indicator function, whose operation rule is: when the logical condition within the parentheses (i.e., ...) is met, the ... The function returns 1 if the condition "contradictory" is true; otherwise, it returns 0. This indicates a logical contradiction label in the output category of the NLI model.

[0097] Step S45, Negotiation Game Trigger Mechanism:

[0098] Real-time monitoring of loss values ​​is required if and only if the criterion is met. Upon establishment, a Multi-Agent Negotiation Protocol is immediately triggered, forcing nodes in a conflict state into a multi-round adversarial dialogue mode until the loss function converges to zero. In other words, conflict information is broadcast to relevant nodes, forcing nodes in a conflict state to adopt the other party's perspective as new context and enter a multi-round adversarial dialogue mode until the loss function converges to zero (i.e., a Nash equilibrium is reached).

[0099] The formatted agent is activated and aggregates the review vectors of all semantic domains only if the global logic check passes.

[0100] In step S5, the weighted fusion specifically adopts a weighted summation model based on confidence decay, the mathematical expression of which is:

[0101]

[0102] in, This indicates the final automated review score for the thesis. This represents the set of all divided semantic review domains, i.e. , This represents the index of the currently traversed field in the collection. Indicates the first The preset base weights for each semantic domain satisfy the normalization constraints. This is used to reflect the differences in the contribution of different chapters to the quality of the paper. Indicates the first Each semantic domain is represented by a raw score given by the corresponding agent. The penalty decay factor is a constant scalar whose value is strictly limited to an open interval. Internally, it is used to control the severity of penalties for uncertainty. Indicates the first The number of self-correction rounds that a domain's review agent undergoes during the critical verification phase. It is a non-negative integer. The larger the value, the more likely that the review comments for that section have undergone multiple rejections and revisions. Therefore, the lower the system's confidence in its initial approval, the lower the success rate of the index item. The contribution to the final score is non-linearly suppressed. This formula means that if a part of the review comments has been rejected and revised multiple times ( The higher the score (the greater the score), the lower the system's confidence in its one-time pass. Therefore, the system performs non-linear suppression on its contribution to the final score, thus ensuring the objectivity and robustness of the scoring. Finally, the system renders and generates a structured PDF review report, including overall evaluation, sub-item scores, and detailed modification suggestions.

[0103] Example 3 is based on the previous two examples, but differs from the previous example in that it also includes a thesis review system, which includes a document reconstruction and routing module, a dynamic graph construction module, a local verification module, a global alignment module, and a result fusion module.

[0104] The document reconstruction and routing module is used to obtain the dissertation documents to be reviewed, parse the documents into structured data objects, and distribute the data blocks in the structured data objects to multiple preset review domains according to the content semantics.

[0105] The dynamic graph construction module is used to extract technical entities from data blocks for at least one core review domain, dynamically divide the technical entities based on the community discovery algorithm to form multiple technical communities, and instantiate at least one review agent for each technical community to construct a parallel review computation graph.

[0106] By dynamically arranging graphs, multiple independent innovations in a paper are adaptively identified, and a dedicated review agent is instantiated for each innovation. This solves the problem of information loss in long text processing and achieves full coverage and fine-grained analysis.

[0107] The local verification module is used to perform first-order local adversarial critique verification on the initial review opinions generated by each review agent, including fact-checking and triggering a self-correction loop based on the confidence level of the fact-checking. A two-order adversarial game architecture is introduced. The first order eliminates the "fact illusion" of the large model through local "execution-critique", and the second order resolves the logical conflicts between multiple agents through the global "mutual implication matrix", which significantly improves the accuracy of the review results.

[0108] The global alignment module is used to perform a second-order global logical consistency game on all review opinions, including determining the logical relationship between each review opinion and triggering multiple rounds of adversarial negotiation until logical convergence when there is a logical conflict.

[0109] The results fusion module is used to weight and fuse the reviews based on their confidence levels to generate the final review results. A score decay algorithm based on correction rounds is employed to reduce the weight of low-confidence review results obtained after multiple corrections, making the final score more robust and scientific, and avoiding the problem of blindly trusting the model's output.

[0110] In this preferred embodiment, the document parsing engine is configured with OCR optical character recognition and layout analysis algorithms to convert unstructured PDF document streams into high-fidelity structured JSON data objects; the agent orchestrator is based on the Kubernetes container orchestration platform and is responsible for parsing dynamic computation graphs, managing the lifecycle of agent containers, and elastically scheduling computing resources; the two-layer critique module has a built-in pre-trained NLI logical reasoning model and a fact consistency verification model, which are used to perform global logical game theory and local fact checking, respectively; and the knowledge enhancement bus encapsulates a high-performance vector database (Vector DB) interface and an Internet instant search (Web Search) API to provide agents with multi-source heterogeneous knowledge context.

[0111] By reconstructing documents and semantic routing, ultra-long text papers are transformed into structured data and distributed to orthogonal review domains. Combined with dynamic graph arrangement technology, comprehensive coverage and fine-grained analysis of the paper content are achieved, effectively solving the problems of distortion in long text processing and bottlenecks in manual review. Furthermore, a two-order adversarial game architecture is introduced, which significantly suppresses the "fact illusion" of large models and eliminates logical conflicts between multi-agent reviewers through a dual verification mechanism of local fact-checking and global logical alignment, greatly improving the accuracy and logical consistency of automated review results. Moreover, a confidence decay algorithm based on self-correcting rounds is used for score fusion, making the review results more objective and robust, providing reliable technical support for the scientific evaluation of dissertations.

[0112] For ease of description, the above system is described by dividing it into various modules based on their functions. Of course, in implementing this invention, the functions of each module can be implemented in one or more software and / or hardware components.

[0113] It should be noted that the method of the present invention can be executed by a single device, such as a computer or server. The method of the present invention can also be applied in a distributed scenario, where multiple devices cooperate to complete the task. In this case, one of these devices may execute only one or more steps of the method of the present invention, and the multiple devices will interact with each other to complete the method.

[0114] It should also be noted that the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0115] The various embodiments in this invention are described in a progressive manner. For the same or similar parts between the various embodiments, please refer to each other. Each embodiment focuses on describing the differences from other embodiments.

[0116] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of one or more embodiments of the invention as described above, which are not provided in detail for the sake of brevity.

[0117] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium may be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium capable of directing a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means implemented in a flow... Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0118] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for reviewing dissertations, characterized in that, include: Step S1: Obtain the thesis document to be reviewed, parse the document into a structured data object, and distribute the data blocks in the structured data object to multiple preset review domains according to the content semantics. Step S2: For at least one core review domain, extract the technical entities from the data block, use the community discovery algorithm to dynamically divide the technical entities and form multiple technical communities, and instantiate at least one review agent for each technical community to construct a parallel review computation graph. Step S3: Perform first-order local adversarial critique verification on the review opinions generated by each review agent. The first-order local adversarial critique verification includes fact checking and triggering a self-correction loop based on the confidence level of the fact checking. Step S4: Perform a second-order global logical consistency game on all review opinions after the first-order local adversarial critique verification. The second-order global logical consistency game includes determining the logical relationship between each review opinion and triggering multiple rounds of adversarial negotiation until logical convergence when there is a logical conflict. Step S5: Weighted fusion is performed based on the confidence level of each review opinion to generate the final review result.

2. The thesis review method as described in claim 1, characterized in that: The multiple preset review domains include the reference domain. Abstract field Method domain Other integrated domains .

3. The thesis review method as described in claim 1, characterized in that: The community detection algorithm is the Louvain algorithm, and the instantiation is containerized instantiation.

4. The thesis review method as described in claim 1, characterized in that: The fact-checking process calculates the confidence level using a fact consistency discriminant function. When the confidence level is lower than a preset threshold, the self-correction loop is triggered.

5. The thesis review method as described in claim 1, characterized in that: The determination of the logical relationship between the review opinions is achieved by constructing a logical mutual implication matrix using a Natural Language Inference (NLI) model.

6. The thesis review method as described in claim 1, characterized in that: The weighted fusion adopts an exponential decay algorithm based on self-correcting rounds; The weight of the review comments is negatively correlated with the number of times they are self-corrected.

7. A thesis review system, wherein the system is used to execute a thesis review method according to any one of claims 1-6, characterized in that, It includes a document reconstruction and routing module, a dynamic graph construction module, a local validation module, a global alignment module, and a result fusion module; The document reconstruction and routing module is used to obtain the thesis documents to be reviewed, parse the documents into structured data objects, and distribute the data blocks in the structured data objects to multiple preset review domains according to the content semantics. The dynamic graph construction module is used to extract technical entities from the data block for at least one core review domain, dynamically divide the technical entities based on the community discovery algorithm to form multiple technical communities, and instantiate at least one review agent for each technical community to construct a parallel review computation graph. The local verification module is used to perform first-order local adversarial critique verification on the initial review opinions generated by each review agent, including fact-checking and triggering a self-correction loop based on the confidence level of the fact-checking. The global alignment module is used to perform a second-order global logical consistency game on all review opinions, including determining the logical relationship between each review opinion and triggering multiple rounds of adversarial negotiation until logical convergence when there is a logical conflict. The result fusion module is used to perform weighted fusion based on the confidence level of each review opinion to generate the final review result.

8. The thesis review system according to claim 7, characterized in that, The document reconstruction and routing module is specifically used to distribute the data block to the reference domain. Abstract field Method domain Other integrated domains .

9. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method according to any one of claims 1-6.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1-6.

Images